Summary
PraisonAI versions before 4.6.62 can bypass token authentication for agent invocation when the advertised PRAISONAI_CALL_AUTH=disabled setting is used. In deployments where agents expose connected tools or private context, unauthenticated users may invoke registered agents and access or affect protected resources.
Affected context
The vulnerability occurs in PraisonAI deployments that expose the /api/v1/agents/{id}/invoke endpoint and configure call authentication as disabled. Under this condition, the token verification logic accepts invocation requests without requiring CALL_SERVER_TOKEN authentication, potentially placing registered agents, connected tools, and private contextual data within reach of unauthenticated network clients.
Exploit preconditions
- Attack surface: The HTTP API endpoint used to invoke registered agents
- Required access: Network access to the exposed PraisonAI API; valid credentials may not be required when call authentication is disabled
- Trigger point: A request to
/api/v1/agents/{id}/invoke - Why exploitation is possible: The authentication verification path treats the configured opt-out as sufficient to accept requests, even when the endpoint is reachable by untrusted clients
Why this happens
The underlying issue is an authentication decision that permits sensitive API operations without an effective identity check when a configuration flag disables call authentication. This maps to CWE-287: Improper Authentication, because the application does not reliably verify that a requester is authorized to invoke an agent. The insecure pattern is allowing a broad configuration-based bypass for a privileged operation without compensating access controls or deployment-scope restrictions.
Potential impact
- Unauthenticated users may invoke registered agents
- Attackers could access information held in agent private context
- Connected tools may be triggered without the intended authorization
- Agent workflows could be manipulated or abused
- Downstream systems and data may be exposed or altered
- Service resources could be consumed through unauthorized requests
General mitigation guidance
Mitigation generally focuses on preserving authentication for sensitive agent operations and reducing the trust placed in configuration-based opt-out settings.
Common defensive patterns include:
- Require strong authentication and authorization for agent invocation endpoints
- Avoid disabling call authentication on interfaces reachable by untrusted networks
- Apply network-level access controls, gateway authentication, or mutually authenticated service communication where appropriate
- Enforce per-agent authorization and least-privilege access to connected tools and private context
- Use secure-by-default configuration profiles that do not expose privileged operations anonymously
- Treat authentication-related configuration changes as security-sensitive and subject them to review and policy enforcement
How teams detect issues like this
Organizations can look for this issue through:
- Configuration scans identifying deployments with call authentication disabled
- API inventories that expose agent invocation endpoints outside trusted network boundaries
- Static analysis and code review checks for authentication branches that accept requests solely because a feature flag disables verification
- Integration tests confirming that unauthenticated requests to sensitive agent operations are rejected under every supported configuration
- Gateway and application logs showing agent invocation without an authenticated principal
- Policy enforcement preventing anonymous access to agent APIs and connected tools
Automated security review tools may help identify authentication bypass patterns and risky configuration combinations.

